diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -384,8 +384,20 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
   if (Subtarget->hasBFE())
     setHasExtractBitsInsn(true);
 
-  setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
-  setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
+  setOperationAction(ISD::FMINNUM, MVT::f32, Custom);
+  setOperationAction(ISD::FMAXNUM, MVT::f32, Custom);
+  setOperationAction(ISD::FMINNUM, MVT::f64, Custom);
+  setOperationAction(ISD::FMAXNUM, MVT::f64, Custom);
+
+
+  // These are really only legal for ieee_mode functions. We should be avoiding
+  // them for functions that don't have ieee_mode enabled, so just say they are
+  // legal.
+  setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal);
+  setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal);
+  setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal);
+  setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal);
+
 
   if (Subtarget->getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
     setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
@@ -474,8 +486,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     // F16 - VOP2 Actions.
     setOperationAction(ISD::BR_CC, MVT::f16, Expand);
     setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
-    setOperationAction(ISD::FMAXNUM, MVT::f16, Legal);
-    setOperationAction(ISD::FMINNUM, MVT::f16, Legal);
+
     setOperationAction(ISD::FDIV, MVT::f16, Custom);
 
     // F16 - VOP3 Actions.
@@ -558,6 +569,17 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     // This isn't really legal, but this avoids the legalizer unrolling it (and
     // allows matching fneg (fabs x) patterns)
     setOperationAction(ISD::FABS, MVT::v2f16, Legal);
+
+    setOperationAction(ISD::FMAXNUM, MVT::f16, Custom);
+    setOperationAction(ISD::FMINNUM, MVT::f16, Custom);
+    setOperationAction(ISD::FMAXNUM_IEEE, MVT::f16, Legal);
+    setOperationAction(ISD::FMINNUM_IEEE, MVT::f16, Legal);
+
+    setOperationAction(ISD::FMINNUM_IEEE, MVT::v4f16, Custom);
+    setOperationAction(ISD::FMAXNUM_IEEE, MVT::v4f16, Custom);
+
+    setOperationAction(ISD::FMINNUM, MVT::v4f16, Expand);
+    setOperationAction(ISD::FMAXNUM, MVT::v4f16, Expand);
   }
 
   if (Subtarget->hasVOP3PInsts()) {
@@ -575,8 +597,10 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::FADD, MVT::v2f16, Legal);
     setOperationAction(ISD::FMUL, MVT::v2f16, Legal);
     setOperationAction(ISD::FMA, MVT::v2f16, Legal);
-    setOperationAction(ISD::FMINNUM, MVT::v2f16, Legal);
-    setOperationAction(ISD::FMAXNUM, MVT::v2f16, Legal);
+
+    setOperationAction(ISD::FMINNUM_IEEE, MVT::v2f16, Legal);
+    setOperationAction(ISD::FMAXNUM_IEEE, MVT::v2f16, Legal);
+
     setOperationAction(ISD::FCANONICALIZE, MVT::v2f16, Legal);
 
     setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i16, Custom);
@@ -596,6 +620,10 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
 
     setOperationAction(ISD::FADD, MVT::v4f16, Custom);
     setOperationAction(ISD::FMUL, MVT::v4f16, Custom);
+
+    setOperationAction(ISD::FMAXNUM, MVT::v2f16, Custom);
+    setOperationAction(ISD::FMINNUM, MVT::v2f16, Custom);
+
     setOperationAction(ISD::FMINNUM, MVT::v4f16, Custom);
     setOperationAction(ISD::FMAXNUM, MVT::v4f16, Custom);
     setOperationAction(ISD::FCANONICALIZE, MVT::v4f16, Custom);
@@ -634,6 +662,8 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
   setTargetDAGCombine(ISD::FSUB);
   setTargetDAGCombine(ISD::FMINNUM);
   setTargetDAGCombine(ISD::FMAXNUM);
+  setTargetDAGCombine(ISD::FMINNUM_IEEE);
+  setTargetDAGCombine(ISD::FMAXNUM_IEEE);
   setTargetDAGCombine(ISD::FMA);
   setTargetDAGCombine(ISD::SMIN);
   setTargetDAGCombine(ISD::SMAX);
@@ -3580,6 +3610,9 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::FNEG:
   case ISD::FCANONICALIZE:
     return splitUnaryVectorOp(Op, DAG);
+  case ISD::FMINNUM:
+  case ISD::FMAXNUM:
+    return lowerFMINNUM_FMAXNUM(Op, DAG);
   case ISD::SHL:
   case ISD::SRA:
   case ISD::SRL:
@@ -3590,10 +3623,10 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::SMAX:
   case ISD::UMIN:
   case ISD::UMAX:
-  case ISD::FMINNUM:
-  case ISD::FMAXNUM:
   case ISD::FADD:
   case ISD::FMUL:
+  case ISD::FMINNUM_IEEE:
+  case ISD::FMAXNUM_IEEE:
     return splitBinaryVectorOp(Op, DAG);
   }
   return SDValue();
@@ -4048,6 +4081,23 @@ SDValue SITargetLowering::lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getNode(ISD::BITCAST, DL, MVT::f16, Trunc);
 }
 
+SDValue SITargetLowering::lowerFMINNUM_FMAXNUM(SDValue Op,
+                                               SelectionDAG &DAG) const {
+  EVT VT = Op.getValueType();
+  bool IsIEEEMode = Subtarget->enableIEEEBit(DAG.getMachineFunction());
+
+  // FIXME: Assert during eslection that this is only selected for
+  // ieee_mode. Currently a combine can produce the ieee version for non-ieee
+  // mode functions, but this happens to be OK since it's only done in cases
+  // where there is known no sNaN.
+  if (IsIEEEMode)
+    return expandFMINNUM_FMAXNUM(Op.getNode(), DAG);
+
+  if (VT == MVT::v4f16)
+    return splitBinaryVectorOp(Op, DAG);
+  return Op;
+}
+
 SDValue SITargetLowering::lowerTRAP(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue Chain = Op.getOperand(0);
@@ -7521,37 +7571,32 @@ bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
 
   case ISD::FMINNUM:
   case ISD::FMAXNUM:
+  case ISD::FMINNUM_IEEE:
+  case ISD::FMAXNUM_IEEE:
   case AMDGPUISD::CLAMP:
   case AMDGPUISD::FMED3:
   case AMDGPUISD::FMAX3:
   case AMDGPUISD::FMIN3: {
     // FIXME: Shouldn't treat the generic operations different based these.
-    bool IsIEEEMode = Subtarget->enableIEEEBit(DAG.getMachineFunction());
-    if (IsIEEEMode) {
-      // snans will be quieted, so we only need to worry about denormals.
-      if (Subtarget->supportsMinMaxDenormModes() ||
-          denormalsEnabledForType(Op.getValueType()))
-        return true;
-
-      // Flushing may be required.
-      // In pre-GFX9 targets V_MIN_F32 and others do not flush denorms. For such
-      // targets need to check their input recursively.
-      return isCanonicalized(DAG, Op.getOperand(0), MaxDepth - 1) &&
-             isCanonicalized(DAG, Op.getOperand(1), MaxDepth - 1);
-    }
+    // However, we aren't really required to flush the result from
+    // minnum/maxnum..
 
+    // snans will be quieted, so we only need to worry about denormals.
     if (Subtarget->supportsMinMaxDenormModes() ||
-        denormalsEnabledForType(Op.getValueType())) {
-      // Only quieting may be necessary.
-      return DAG.isKnownNeverSNaN(Op.getOperand(0)) &&
-             DAG.isKnownNeverSNaN(Op.getOperand(1));
+        denormalsEnabledForType(Op.getValueType()))
+      return true;
+
+    // Flushing may be required.
+    // In pre-GFX9 targets V_MIN_F32 and others do not flush denorms. For such
+    // targets need to check their input recursively.
+
+    // FIXME: Does this apply with clamp? It's implemented with max.
+    for (unsigned I = 0, E = Op.getNumOperands(); I != E; ++I) {
+      if (!isCanonicalized(DAG, Op.getOperand(I), MaxDepth - 1))
+        return false;
     }
 
-    // Flushing and quieting may be necessary
-    // With ieee_mode off, the nan is returned as-is, so if it is an sNaN it
-    // needs to be quieted.
-    return isCanonicalized(DAG, Op.getOperand(0), MaxDepth - 1) &&
-           isCanonicalized(DAG, Op.getOperand(1), MaxDepth - 1);
+    return true;
   }
   case ISD::SELECT: {
     return isCanonicalized(DAG, Op.getOperand(1), MaxDepth - 1) &&
@@ -7578,6 +7623,21 @@ bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
     // Could be anything.
     return false;
 
+  case ISD::BITCAST: {
+    // Hack round the mess we make when legalizing extract_vector_elt
+    SDValue Src = Op.getOperand(0);
+    if (Src.getValueType() == MVT::i16 &&
+        Src.getOpcode() == ISD::TRUNCATE) {
+      SDValue TruncSrc = Src.getOperand(0);
+      if (TruncSrc.getValueType() == MVT::i32 &&
+          TruncSrc.getOpcode() == ISD::BITCAST &&
+          TruncSrc.getOperand(0).getValueType() == MVT::v2f16) {
+        return isCanonicalized(DAG, TruncSrc.getOperand(0), MaxDepth - 1);
+      }
+    }
+
+    return false;
+  }
   case ISD::INTRINSIC_WO_CHAIN: {
     unsigned IntrinsicID
       = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
@@ -7603,7 +7663,6 @@ bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
 }
 
 // Constant fold canonicalize.
-
 SDValue SITargetLowering::getCanonicalConstantFP(
   SelectionDAG &DAG, const SDLoc &SL, EVT VT, const APFloat &C) const {
   // Flush denormals to 0 if not enabled.
@@ -7699,18 +7758,40 @@ SDValue SITargetLowering::performFCanonicalizeCombine(
     }
   }
 
+  unsigned SrcOpc = N0.getOpcode();
+
+  // If it's free to do so, push canonicalizes further up the source, which may
+  // find a canonical source.
+  //
+  // TODO: More opcodes. Note this is unsafe for the the _ieee minnum/maxnum for
+  // sNaNs.
+  if (SrcOpc == ISD::FMINNUM || SrcOpc == ISD::FMAXNUM) {
+    auto *CRHS = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
+    if (CRHS && N0.hasOneUse()) {
+      SDLoc SL(N);
+      SDValue Canon0 = DAG.getNode(ISD::FCANONICALIZE, SL, VT,
+                                   N0.getOperand(0));
+      SDValue Canon1 = getCanonicalConstantFP(DAG, SL, VT, CRHS->getValueAPF());
+      DCI.AddToWorklist(Canon0.getNode());
+
+      return DAG.getNode(N0.getOpcode(), SL, VT, Canon0, Canon1);
+    }
+  }
+
   return isCanonicalized(DAG, N0) ? N0 : SDValue();
 }
 
 static unsigned minMaxOpcToMin3Max3Opc(unsigned Opc) {
   switch (Opc) {
   case ISD::FMAXNUM:
+  case ISD::FMAXNUM_IEEE:
     return AMDGPUISD::FMAX3;
   case ISD::SMAX:
     return AMDGPUISD::SMAX3;
   case ISD::UMAX:
     return AMDGPUISD::UMAX3;
   case ISD::FMINNUM:
+  case ISD::FMINNUM_IEEE:
     return AMDGPUISD::FMIN3;
   case ISD::SMIN:
     return AMDGPUISD::SMIN3;
@@ -7877,6 +7958,7 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
 
   // fminnum(fmaxnum(x, K0), K1), K0 < K1 && !is_snan(x) -> fmed3(x, K0, K1)
   if (((Opc == ISD::FMINNUM && Op0.getOpcode() == ISD::FMAXNUM) ||
+       (Opc == ISD::FMINNUM_IEEE && Op0.getOpcode() == ISD::FMAXNUM_IEEE) ||
        (Opc == AMDGPUISD::FMIN_LEGACY &&
         Op0.getOpcode() == AMDGPUISD::FMAX_LEGACY)) &&
       (VT == MVT::f32 || VT == MVT::f64 ||
@@ -7995,7 +8077,9 @@ SDValue SITargetLowering::performExtractVectorEltCombine(
     case ISD::SMIN:
     case ISD::SMAX:
     case ISD::FMAXNUM:
-    case ISD::FMINNUM: {
+    case ISD::FMINNUM:
+    case ISD::FMAXNUM_IEEE:
+    case ISD::FMINNUM_IEEE: {
       SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
                                  Vec.getOperand(0), Idx);
       SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
@@ -8595,13 +8679,15 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
     return performSetCCCombine(N, DCI);
   case ISD::FMAXNUM:
   case ISD::FMINNUM:
+  case ISD::FMAXNUM_IEEE:
+  case ISD::FMINNUM_IEEE:
   case ISD::SMAX:
   case ISD::SMIN:
   case ISD::UMAX:
   case ISD::UMIN:
   case AMDGPUISD::FMIN_LEGACY:
   case AMDGPUISD::FMAX_LEGACY: {
-    if (DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
+    if (//DCI.getDAGCombineLevel() >= AfterLegalizeDAG &&
         getTargetMachine().getOptLevel() > CodeGenOpt::None)
       return performMinMaxCombine(N, DCI);
     break;
@@ -9320,3 +9406,17 @@ bool SITargetLowering::denormalsEnabledForType(EVT VT) const {
     return false;
   }
 }
+
+bool SITargetLowering::isKnownNeverNaNForTargetNode(SDValue Op,
+                                                    const SelectionDAG &DAG,
+                                                    bool SNaN,
+                                                    unsigned Depth) const {
+  if (Op.getOpcode() == AMDGPUISD::CLAMP) {
+    if (Subtarget->enableDX10Clamp())
+      return true; // Clamped to 0.
+    return DAG.isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1);
+  }
+
+  return AMDGPUTargetLowering::isKnownNeverNaNForTargetNode(Op, DAG,
+                                                            SNaN, Depth);
+}
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -110,6 +110,7 @@ class SITargetLowering final : public AMDGPUTargetLowering {
 
   /// Custom lowering for ISD::FP_ROUND for MVT::f16.
   SDValue lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerFMINNUM_FMAXNUM(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue getSegmentAperture(unsigned AS, const SDLoc &DL,
                              SelectionDAG &DAG) const;
@@ -346,6 +347,11 @@ class SITargetLowering final : public AMDGPUTargetLowering {
   bool isCanonicalized(SelectionDAG &DAG, SDValue Op,
                        unsigned MaxDepth = 5) const;
   bool denormalsEnabledForType(EVT VT) const;
+
+  bool isKnownNeverNaNForTargetNode(SDValue Op,
+                                    const SelectionDAG &DAG,
+                                    bool SNaN = false,
+                                    unsigned Depth = 0) const override;
 };
 
 } // End namespace llvm

diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1645,21 +1645,22 @@ def : IntMed3Pat<V_MED3_U32, umax, umax_oneuse, umin_oneuse>;
 // This matches 16 permutations of
 // max(min(x, y), min(max(x, y), z))
 class FPMed3Pat<ValueType vt,
+                //SDPatternOperator max, SDPatternOperator min,
                 Instruction med3Inst> : GCNPat<
-  (fmaxnum (fminnum_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
+  (fmaxnum_like (fminnum_like_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
                            (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
-           (fminnum_oneuse (fmaxnum_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
+           (fminnum_like_oneuse (fmaxnum_like_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
                                            (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
                            (vt (VOP3Mods_nnan vt:$src2, i32:$src2_mods)))),
   (med3Inst $src0_mods, $src0, $src1_mods, $src1, $src2_mods, $src2, DSTCLAMP.NONE, DSTOMOD.NONE)
 >;
 
 class FP16Med3Pat<ValueType vt,
                 Instruction med3Inst> : GCNPat<
-  (fmaxnum (fminnum_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
-                           (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
-           (fminnum_oneuse (fmaxnum_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
-                                           (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
+  (fmaxnum_like (fminnum_like_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
+                                     (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
+           (fminnum_like_oneuse (fmaxnum_like_oneuse (VOP3Mods_nnan vt:$src0, i32:$src0_mods),
+                                                     (VOP3Mods_nnan vt:$src1, i32:$src1_mods)),
                            (vt (VOP3Mods_nnan vt:$src2, i32:$src2_mods)))),
   (med3Inst $src0_mods, $src0, $src1_mods, $src1, $src2_mods, $src2, DSTCLAMP.NONE)
 >;

diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -393,8 +393,8 @@ defm V_MUL_I32_I24 : VOP2Inst <"v_mul_i32_i24", VOP_PAT_GEN<VOP_I32_I32_I32, 2>,
 defm V_MUL_HI_I32_I24 : VOP2Inst <"v_mul_hi_i32_i24", VOP_PAT_GEN<VOP_I32_I32_I32, 2>, AMDGPUmulhi_i24>;
 defm V_MUL_U32_U24 : VOP2Inst <"v_mul_u32_u24", VOP_PAT_GEN<VOP_I32_I32_I32, 2>, AMDGPUmul_u24>;
 defm V_MUL_HI_U32_U24 : VOP2Inst <"v_mul_hi_u32_u24", VOP_PAT_GEN<VOP_I32_I32_I32, 2>, AMDGPUmulhi_u24>;
-defm V_MIN_F32 : VOP2Inst <"v_min_f32", VOP_F32_F32_F32, fminnum>;
-defm V_MAX_F32 : VOP2Inst <"v_max_f32", VOP_F32_F32_F32, fmaxnum>;
+defm V_MIN_F32 : VOP2Inst <"v_min_f32", VOP_F32_F32_F32, fminnum_like>;
+defm V_MAX_F32 : VOP2Inst <"v_max_f32", VOP_F32_F32_F32, fmaxnum_like>;
 defm V_MIN_I32 : VOP2Inst <"v_min_i32", VOP_PAT_GEN<VOP_I32_I32_I32>, smin>;
 defm V_MAX_I32 : VOP2Inst <"v_max_i32", VOP_PAT_GEN<VOP_I32_I32_I32>, smax>;
 defm V_MIN_U32 : VOP2Inst <"v_min_u32", VOP_PAT_GEN<VOP_I32_I32_I32>, umin>;
@@ -556,8 +556,8 @@ defm V_ADD_U16 : VOP2Inst <"v_add_u16", VOP_I16_I16_I16>;
 defm V_SUB_U16 : VOP2Inst <"v_sub_u16" , VOP_I16_I16_I16>;
 defm V_SUBREV_U16 : VOP2Inst <"v_subrev_u16", VOP_I16_I16_I16, null_frag, "v_sub_u16">;
 defm V_MUL_LO_U16 : VOP2Inst <"v_mul_lo_u16", VOP_I16_I16_I16>;
-defm V_MAX_F16 : VOP2Inst <"v_max_f16", VOP_F16_F16_F16, fmaxnum>;
-defm V_MIN_F16 : VOP2Inst <"v_min_f16", VOP_F16_F16_F16, fminnum>;
+defm V_MAX_F16 : VOP2Inst <"v_max_f16", VOP_F16_F16_F16, fmaxnum_like>;
+defm V_MIN_F16 : VOP2Inst <"v_min_f16", VOP_F16_F16_F16, fminnum_like>;
 defm V_MAX_U16 : VOP2Inst <"v_max_u16", VOP_I16_I16_I16>;
 defm V_MAX_I16 : VOP2Inst <"v_max_i16", VOP_I16_I16_I16>;
 defm V_MIN_U16 : VOP2Inst <"v_min_u16", VOP_I16_I16_I16>;

diff --git a/llvm/lib/Target/AMDGPU/VOP3Instructions.td b/llvm/lib/Target/AMDGPU/VOP3Instructions.td
@@ -295,8 +295,8 @@ let SchedRW = [WriteDoubleAdd] in {
 def V_FMA_F64 : VOP3Inst <"v_fma_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, fma>;
 def V_ADD_F64 : VOP3Inst <"v_add_f64", VOP3_Profile<VOP_F64_F64_F64>, fadd, 1>;
 def V_MUL_F64 : VOP3Inst <"v_mul_f64", VOP3_Profile<VOP_F64_F64_F64>, fmul, 1>;
-def V_MIN_F64 : VOP3Inst <"v_min_f64", VOP3_Profile<VOP_F64_F64_F64>, fminnum, 1>;
-def V_MAX_F64 : VOP3Inst <"v_max_f64", VOP3_Profile<VOP_F64_F64_F64>, fmaxnum, 1>;
+def V_MIN_F64 : VOP3Inst <"v_min_f64", VOP3_Profile<VOP_F64_F64_F64>, fminnum_like, 1>;
+def V_MAX_F64 : VOP3Inst <"v_max_f64", VOP3_Profile<VOP_F64_F64_F64>, fmaxnum_like, 1>;
 } // End SchedRW = [WriteDoubleAdd]
 
 let SchedRW = [WriteQuarterRate32] in {

diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -48,8 +48,8 @@ def V_PK_MAD_U16 : VOP3PInst<"v_pk_mad_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16_
 
 def V_PK_ADD_F16 : VOP3PInst<"v_pk_add_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fadd>;
 def V_PK_MUL_F16 : VOP3PInst<"v_pk_mul_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmul>;
-def V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum>;
-def V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fminnum>;
+def V_PK_MAX_F16 : VOP3PInst<"v_pk_max_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fmaxnum_like>;
+def V_PK_MIN_F16 : VOP3PInst<"v_pk_min_f16", VOP3_Profile<VOP_V2F16_V2F16_V2F16>, fminnum_like>;
 
 def V_PK_ADD_U16 : VOP3PInst<"v_pk_add_u16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>, add>;
 def V_PK_ADD_I16 : VOP3PInst<"v_pk_add_i16", VOP3_Profile<VOP_V2I16_V2I16_V2I16>>;

diff --git a/llvm/test/CodeGen/AMDGPU/clamp.ll b/llvm/test/CodeGen/AMDGPU/clamp.ll
@@ -74,7 +74,8 @@ define amdgpu_kernel void @v_clamp_negzero_f32(float addrspace(1)* %out, float a
 
 ; GCN-LABEL: {{^}}v_clamp_negzero_maybe_snan_f32:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
-; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0x80000000, [[A]]
+; GCN: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[A]]
+; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0x80000000, [[QUIET]]
 ; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], 1.0, [[MAX]]
 define amdgpu_kernel void @v_clamp_negzero_maybe_snan_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
@@ -90,8 +91,17 @@ define amdgpu_kernel void @v_clamp_negzero_maybe_snan_f32(float addrspace(1)* %o
 
 ; GCN-LABEL: {{^}}v_clamp_multi_use_max_f32:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
-; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0, [[A]]
-; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], 1.0, [[MAX]]
+; GCN: v_mul_f32_e32 [[QUIET_A:v[0-9]+]], 1.0, [[A]]
+; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0, [[QUIET_A]]
+; GCN: v_min_f32_e32 [[MED:v[0-9]+]], 1.0, [[QUIET_A]]
+; GCN-NOT: [[MAX]]
+; GCN-NOT: [[MED]]
+
+; SI: buffer_store_dword [[MED]]
+; SI: buffer_store_dword [[MAX]]
+
+; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[MED]]
+; GFX89: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[MAX]]
 define amdgpu_kernel void @v_clamp_multi_use_max_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
@@ -406,8 +416,8 @@ define amdgpu_kernel void @v_clamp_f32_snan_dx10clamp(float addrspace(1)* %out,
 
 ; GCN-LABEL: {{^}}v_clamp_f32_snan_no_dx10clamp:
 ; GCN: {{buffer|flat|global}}_load_dword [[A:v[0-9]+]]
-; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0, [[A]]
-; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], 1.0, [[MAX]]
+; GCN: v_mul_f32_e32 [[QUIET_A:v[0-9]+]], 1.0, [[A]]
+; GCN: v_med3_f32 {{v[0-9]+}}, [[QUIET_A]], 0, 1.0
 define amdgpu_kernel void @v_clamp_f32_snan_no_dx10clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #4 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid

diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
@@ -455,14 +455,13 @@ define amdgpu_kernel void @test_fold_canonicalize_qNaN_value_f32(float addrspace
 }
 
 ; GCN-LABEL: test_fold_canonicalize_minnum_value_from_load_f32_ieee_mode:
-; GCN: v_min_f32_e32 [[V:v[0-9]+]], 0, v{{[0-9]+}}
-; GFX9-NOT: v_max
-; GFX9-NOT: v_mul
-
-; VI-DENORM-NOT: v_max_f32
-; VI-DENORM-NOT: v_mul_f32
+; GCN: {{flat|global}}_load_dword [[VAL:v[0-9]+]]
+; GCN-FLUSH: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[VAL]]
+; GCN-DENORM: v_max_f32_e32 [[QUIET:v[0-9]+]], [[VAL]], [[VAL]]
+; GCN: v_min_f32_e32 [[V:v[0-9]+]], 0, [[QUIET]]
 
-; VI-FLUSH: v_mul_f32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
+; GCN-NOT: v_max
+; GCN-NOT: v_mul
 
 ; GFX9: {{flat|global}}_store_dword v[{{[0-9:]+}}], [[V]]
 define amdgpu_kernel void @test_fold_canonicalize_minnum_value_from_load_f32_ieee_mode(float addrspace(1)* %arg) {
@@ -476,15 +475,13 @@ define amdgpu_kernel void @test_fold_canonicalize_minnum_value_from_load_f32_iee
 }
 
 ; GCN-LABEL: test_fold_canonicalize_minnum_value_from_load_f32_nnan_ieee_mode:
-; GCN: v_min_f32_e32 v{{[0-9]+}}, 0, v{{[0-9]+}}
-
-; GFX9-NOT: v_max
-; GFX9-NOT: v_mul
-
-
-; VI-DENORM-NOT: v_max
-; VI-DENORM-NOT: v_mul
 ; VI-FLUSH: v_mul_f32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}
+; GCN-DENORM-NOT: v_max
+; GCN-DENORM-NOT: v_mul
+
+; GCN: v_min_f32_e32 v{{[0-9]+}}, 0, v{{[0-9]+}}
+; GCN-DENORM-NOT: v_max
+; GCN-DENORM-NOT: v_mul
 
 ; GFX9: {{flat|global}}_store_dword v[{{[0-9:]+}}]
 define amdgpu_kernel void @test_fold_canonicalize_minnum_value_from_load_f32_nnan_ieee_mode(float addrspace(1)* %arg) #1 {
@@ -530,13 +527,19 @@ define amdgpu_kernel void @test_fold_canonicalize_sNaN_value_f32(float addrspace
 }
 
 ; GCN-LABEL: test_fold_canonicalize_denorm_value_f32:
-; GFX9:  v_min_f32_e32 [[RESULT:v[0-9]+]], 0x7fffff, v{{[0-9]+}}
+; GCN: {{flat|global}}_load_dword [[VAL:v[0-9]+]]
+
+; GFX9-DENORM: v_max_f32_e32 [[QUIET:v[0-9]+]], [[VAL]], [[VAL]]
+; GFX9-DENORM: v_min_f32_e32 [[RESULT:v[0-9]+]], 0x7fffff, [[QUIET]]
 
-; VI-FLUSH: v_min_f32_e32 [[V0:v[0-9]+]], 0x7fffff, v{{[0-9]+}}
-; VI-FLUSH: v_mul_f32_e32 [[RESULT:v[0-9]+]], 1.0, [[V0]]
+; GFX9-FLUSH: v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[VAL]]
+; GFX9-FLUSH: v_min_f32_e32 [[RESULT:v[0-9]+]], 0, [[QUIET]]
 
-; VI-DENORM: v_min_f32_e32 [[RESULT:v[0-9]+]], 0x7fffff, v{{[0-9]+}}
 
+; VI-FLUSH: v_mul_f32_e32 [[QUIET_V0:v[0-9]+]], 1.0, [[VAL]]
+; VI-FLUSH: v_min_f32_e32 [[RESULT:v[0-9]+]], 0, [[QUIET_V0]]
+
+; VI-DENORM: v_min_f32_e32 [[RESULT:v[0-9]+]], 0x7fffff, [[VAL]]
 
 ; GCN-NOT: v_mul
 ; GCN-NOT: v_max
@@ -552,11 +555,14 @@ define amdgpu_kernel void @test_fold_canonicalize_denorm_value_f32(float addrspa
 }
 
 ; GCN-LABEL: test_fold_canonicalize_maxnum_value_from_load_f32_ieee_mode:
-; GFX9:  v_max_f32_e32 [[RESULT:v[0-9]+]], 0, v{{[0-9]+}}
-; VI-FLUSH:    v_max_f32_e32 [[V0:v[0-9]+]], 0, v{{[0-9]+}}
-; VI-FLUSH:    v_mul_f32_e32 [[RESULT:v[0-9]+]], 1.0, [[V0]]
+; GCN: {{flat|global}}_load_dword [[VAL:v[0-9]+]]
+
+; GFX9:  v_max_f32_e32 [[RESULT:v[0-9]+]], 0, [[VAL]]
+
+; VI-FLUSH:    v_mul_f32_e32 [[QUIET:v[0-9]+]], 1.0, [[VAL]]
+; VI-FLUSH:    v_max_f32_e32 [[RESULT:v[0-9]+]], 0, [[QUIET]]
 
-; VI-DENORM: v_max_f32_e32 [[RESULT:v[0-9]+]], 0, v{{[0-9]+}}
+; VI-DENORM: v_max_f32_e32 [[RESULT:v[0-9]+]], 0, [[VAL]]
 
 ; GCN-NOT: v_mul
 ; GCN-NOT: v_max
@@ -707,16 +713,21 @@ define amdgpu_kernel void @test_fold_canonicalize_select_value_f32(float addrspa
 
 ; Need to quiet the nan with a separate instruction since it will be
 ; passed through the minnum.
+; FIXME: canonicalize doens't work correctly without ieee_mode
 
 ; GCN-LABEL: {{^}}test_fold_canonicalize_minnum_value_no_ieee_mode:
+; GFX9-NOT: v0
+; GFX9-NOT: v1
 ; GFX9: v_min_f32_e32 v0, v0, v1
-; GFX9-FLUSH-NEXT: v_mul_f32_e32 v0, 1.0, v0
-; GFX9-DENORM-NEXT: v_max_f32_e32 v0, v0, v0
 ; GFX9-NEXT: ; return to shader
 
-; VI: v_min_f32_e32 v0, v0, v1
-; VI-FLUSH: v_mul_f32_e32 v0, 1.0, v0
-; VI-DENORM: v_max_f32_e32 v0, v0, v0
+; VI-FLUSH: v_min_f32_e32 v0, v0, v1
+; VI-FLUSH-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; VI-FLUSH-NEXT: ; return
+
+; VI-DENORM-NOT: v0
+; VI-DENORM: v_min_f32_e32 v0, v0, v1
+; VI-DENORM-NEXT: ; return
 define amdgpu_ps float @test_fold_canonicalize_minnum_value_no_ieee_mode(float %arg0, float %arg1) {
   %v = tail call float @llvm.minnum.f32(float %arg0, float %arg1)
   %canonicalized = tail call float @llvm.canonicalize.f32(float %v)
@@ -727,8 +738,14 @@ define amdgpu_ps float @test_fold_canonicalize_minnum_value_no_ieee_mode(float %
 ; GFX9: v_min_f32_e32 v0, v0, v1
 ; GFX9-NEXT: s_setpc_b64
 
-; VI: v_min_f32_e32 v0, v0, v1
-; VI-FLUSH-NEXT: v_mul_f32_e32 v0, 1.0, v0
+; VI-FLUSH-DAG: v_mul_f32_e32 v0, 1.0, v0
+; VI-FLUSH-DAG: v_mul_f32_e32 v1, 1.0, v1
+; VI-FLUSH: v_min_f32_e32 v0, v0, v1
+
+; VI-DENORM-DAG: v_max_f32_e32 v0, v0, v0
+; VI-DENORM-DAG: v_max_f32_e32 v1, v1, v1
+; VI-DENORM: v_min_f32_e32 v0, v0, v1
+
 ; VI-NEXT: s_setpc_b64
 define float @test_fold_canonicalize_minnum_value_ieee_mode(float %arg0, float %arg1) {
   %v = tail call float @llvm.minnum.f32(float %arg0, float %arg1)

diff --git a/llvm/test/CodeGen/AMDGPU/fmax3.f64.ll b/llvm/test/CodeGen/AMDGPU/fmax3.f64.ll
@@ -4,11 +4,14 @@
 declare double @llvm.maxnum.f64(double, double) nounwind readnone
 
 ; SI-LABEL: {{^}}test_fmax3_f64:
-; SI-DAG: buffer_load_dwordx2 [[REGA:v\[[0-9]+:[0-9]+\]]], off, s[{{[0-9]+:[0-9]+}}], 0{{$}}
-; SI-DAG: buffer_load_dwordx2 [[REGB:v\[[0-9]+:[0-9]+\]]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:8
-; SI: v_max_f64 [[REGA]], [[REGA]], [[REGB]]
+; SI: buffer_load_dwordx2 [[REGA:v\[[0-9]+:[0-9]+\]]], off, s[{{[0-9]+:[0-9]+}}], 0{{$}}
+; SI: buffer_load_dwordx2 [[REGB:v\[[0-9]+:[0-9]+\]]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:8
 ; SI: buffer_load_dwordx2 [[REGC:v\[[0-9]+:[0-9]+\]]], off, s[{{[0-9]+:[0-9]+}}], 0 offset:16
-; SI: v_max_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[REGA]], [[REGC]]
+; SI: v_max_f64 [[QUIET_A:v\[[0-9]+:[0-9]+\]]], [[REGA]], [[REGA]]
+; SI: v_max_f64 [[QUIET_B:v\[[0-9]+:[0-9]+\]]], [[REGB]], [[REGB]]
+; SI: v_max_f64 [[MAX0:v\[[0-9]+:[0-9]+\]]], [[QUIET_A]], [[QUIET_B]]
+; SI: v_max_f64 [[QUIET_C:v\[[0-9]+:[0-9]+\]]], [[REGC]], [[REGC]]
+; SI: v_max_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[MAX0]], [[QUIET_C]]
 ; SI: buffer_store_dwordx2 [[RESULT]],
 ; SI: s_endpgm
 define amdgpu_kernel void @test_fmax3_f64(double addrspace(1)* %out, double addrspace(1)* %aptr) nounwind {

diff --git a/llvm/test/CodeGen/AMDGPU/fmax3.ll b/llvm/test/CodeGen/AMDGPU/fmax3.ll
@@ -48,8 +48,11 @@ define amdgpu_kernel void @test_fmax3_olt_1_f32(float addrspace(1)* %out, float
 ; SI: v_max3_f32 [[RESULT_F32:v[0-9]+]], [[CVT_A]], [[CVT_B]], [[CVT_C]]
 ; SI: v_cvt_f16_f32_e32 [[RESULT:v[0-9]+]], [[RESULT_F32]]
 
-; VI: v_max_f16_e32
-; VI: v_max_f16_e32 [[RESULT:v[0-9]+]],
+; VI-DAG: v_max_f16_e32 [[QUIET_A:v[0-9]+]], [[REGA]], [[REGA]]
+; VI-DAG: v_max_f16_e32 [[QUIET_B:v[0-9]+]], [[REGB]], [[REGB]]
+; VI: v_max_f16_e32 [[MAX0:v[0-9]+]], [[QUIET_A]], [[QUIET_B]]
+; VI: v_max_f16_e32 [[QUIET_C:v[0-9]+]], [[REGC]], [[REGC]]
+; VI: v_max_f16_e32 [[RESULT:v[0-9]+]], [[MAX0]], [[QUIET_C]]
 
 ; GFX9: v_max3_f16 [[RESULT:v[0-9]+]], [[REGA]], [[REGB]], [[REGC]]
 ; GCN: buffer_store_short [[RESULT]],
@@ -75,8 +78,11 @@ define amdgpu_kernel void @test_fmax3_olt_0_f16(half addrspace(1)* %out, half ad
 ; SI: v_max3_f32 [[RESULT_F32:v[0-9]+]], [[CVT_C]], [[CVT_A]], [[CVT_B]]
 ; SI: v_cvt_f16_f32_e32 [[RESULT:v[0-9]+]], [[RESULT_F32]]
 
-; VI: v_max_f16_e32
-; VI: v_max_f16_e32 [[RESULT:v[0-9]+]],
+; VI-DAG: v_max_f16_e32 [[QUIET_A:v[0-9]+]], [[REGA]], [[REGA]]
+; VI-DAG: v_max_f16_e32 [[QUIET_B:v[0-9]+]], [[REGB]], [[REGB]]
+; VI: v_max_f16_e32 [[MAX0:v[0-9]+]], [[QUIET_A]], [[QUIET_B]]
+; VI: v_max_f16_e32 [[QUIET_C:v[0-9]+]], [[REGC]], [[REGC]]
+; VI: v_max_f16_e32 [[RESULT:v[0-9]+]], [[QUIET_C]], [[MAX0]]
 
 ; GFX9: v_max3_f16 [[RESULT:v[0-9]+]], [[REGC]], [[REGA]], [[REGB]]
 ; GCN: buffer_store_short [[RESULT]],
@@ -100,22 +106,25 @@ define amdgpu_kernel void @test_fmax3_olt_1_f16(half addrspace(1)* %out, half ad
 ; SI-NEXT: v_max3_f32
 ; SI-NEXT: v_max3_f32
 
-; VI: v_max_f16_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; VI: v_max_f16_e32 v0, v0, v1
-; VI: v_max_f16_sdwa v1, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI: v_max_f16_e32 v0, v2, v0
-; VI: v_max_f16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; VI: v_max_f16_e32 v0, v0, v3
-; VI: v_or_b32_e32 v0, v0, v1
-
-; GFX9: v_pk_max_f16
+; VI: s_waitcnt
+; VI-NEXT: v_max_f16_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-NEXT: v_max_f16_e32 v0, v0, v1
+; VI-NEXT: v_max_f16_sdwa v1, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-NEXT: v_max_f16_e32 v0, v2, v0
+; VI-NEXT: v_max_f16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; VI-NEXT: v_max_f16_e32 v0, v0, v3
+; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: s_setpc_b64
+
+; GFX9: s_waitcnt
+; GFX9-NEXT: v_pk_max_f16
 ; GFX9-NEXT: v_pk_max_f16
 ; GFX9-NEXT: v_pk_max_f16
-define <2 x half> @no_fmax3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) {
+define <2 x half> @no_fmax3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) #2 {
 entry:
-  %max = tail call fast <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b)
-  %max1 = tail call fast <2 x half> @llvm.maxnum.v2f16(<2 x half> %c, <2 x half> %max)
-  %res = tail call fast <2 x half> @llvm.maxnum.v2f16(<2 x half> %max1, <2 x half> %d)
+  %max = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b)
+  %max1 = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %c, <2 x half> %max)
+  %res = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %max1, <2 x half> %d)
   ret <2 x half> %res
 }
 
@@ -126,3 +135,4 @@ declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>)
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone speculatable }
+attributes #2 = { nounwind "no-nans-fp-math"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll b/llvm/test/CodeGen/AMDGPU/fmax_legacy.f16.ll
@@ -97,7 +97,7 @@ define <2 x half> @test_fmax_legacy_ugt_v2f16(<2 x half> %a, <2 x half> %b) #0 {
 ; VI-NNAN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; VI-NNAN-NEXT:    v_max_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_max_f16_e32 v0, v0, v1
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v2
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmax_legacy_ugt_v2f16:
@@ -178,7 +178,7 @@ define <3 x half> @test_fmax_legacy_ugt_v3f16(<3 x half> %a, <3 x half> %b) #0 {
 ; VI-NNAN-NEXT:    v_max_f16_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_max_f16_e32 v0, v0, v2
 ; VI-NNAN-NEXT:    v_max_f16_e32 v1, v1, v3
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v4
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmax_legacy_ugt_v3f16:
@@ -283,8 +283,8 @@ define <4 x half> @test_fmax_legacy_ugt_v4f16(<4 x half> %a, <4 x half> %b) #0 {
 ; VI-NNAN-NEXT:    v_max_f16_e32 v1, v1, v3
 ; VI-NNAN-NEXT:    v_max_f16_sdwa v5, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_max_f16_e32 v0, v0, v2
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v5
-; VI-NNAN-NEXT:    v_or_b32_e32 v1, v1, v4
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v1, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmax_legacy_ugt_v4f16:
@@ -437,10 +437,10 @@ define <8 x half> @test_fmax_legacy_ugt_v8f16(<8 x half> %a, <8 x half> %b) #0 {
 ; VI-NNAN-NEXT:    v_max_f16_e32 v1, v1, v5
 ; VI-NNAN-NEXT:    v_max_f16_sdwa v11, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_max_f16_e32 v0, v0, v4
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v11
-; VI-NNAN-NEXT:    v_or_b32_e32 v1, v1, v10
-; VI-NNAN-NEXT:    v_or_b32_e32 v2, v2, v9
-; VI-NNAN-NEXT:    v_or_b32_e32 v3, v3, v8
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v1, v1, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v2, v2, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v3, v3, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmax_legacy_ugt_v8f16:

diff --git a/llvm/test/CodeGen/AMDGPU/fmax_legacy.ll b/llvm/test/CodeGen/AMDGPU/fmax_legacy.ll
@@ -1,13 +1,22 @@
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GCN-SAFE -check-prefix=FUNC %s
-; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN-NONAN -check-prefix=GCN -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN-SAFE,SI-SAFE,GCN,FUNC %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI-NONAN,GCN-NONAN,GCN,FUNC %s
+
+; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI-SAFE,GCN-SAFE,GCN,FUNC %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI-NONAN,GCN-NONAN,GCN,FUNC %s
+
 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -enable-var-scope -check-prefix=EG -check-prefix=FUNC %s
 
 declare i32 @llvm.r600.read.tidig.x() #1
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_uge_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+
+; VI-SAFE: v_cmp_nlt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_max_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 
 ; EG: MAX
@@ -26,12 +35,16 @@ define amdgpu_kernel void @test_fmax_legacy_uge_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_uge_f32_nnan_src:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
 ; GCN-DAG: v_add_f32_e32 [[ADD_A:v[0-9]+]], 1.0, [[A]]
 ; GCN-DAG: v_add_f32_e32 [[ADD_B:v[0-9]+]], 2.0, [[B]]
 
-; GCN-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[ADD_B]], [[ADD_A]]
+; SI-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[ADD_B]], [[ADD_A]]
+
+; VI-SAFE: v_cmp_nlt_f32_e32 vcc, [[ADD_A]], [[ADD_B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[ADD_B]], [[ADD_A]]
+
 ; GCN-NONAN: v_max_f32_e32 {{v[0-9]+}}, [[ADD_A]], [[ADD_B]]
 
 ; EG: MAX
@@ -52,9 +65,14 @@ define amdgpu_kernel void @test_fmax_legacy_uge_f32_nnan_src(float addrspace(1)*
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_oge_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+
+; VI-SAFE: v_cmp_ge_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_max_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 ; EG: MAX
 define amdgpu_kernel void @test_fmax_legacy_oge_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
@@ -72,9 +90,15 @@ define amdgpu_kernel void @test_fmax_legacy_oge_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_ugt_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+
+; VI-SAFE: v_cmp_nle_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
+
 ; GCN-NONAN: v_max_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 ; EG: MAX
 define amdgpu_kernel void @test_fmax_legacy_ugt_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
@@ -92,9 +116,14 @@ define amdgpu_kernel void @test_fmax_legacy_ugt_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_ogt_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+
+; VI-SAFE: v_cmp_gt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_max_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 ; EG: MAX
 define amdgpu_kernel void @test_fmax_legacy_ogt_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
@@ -112,9 +141,15 @@ define amdgpu_kernel void @test_fmax_legacy_ogt_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_ogt_v1f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_max_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+
+; VI-SAFE: v_cmp_gt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
+
 ; GCN-NONAN: v_max_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 ; EG: MAX
 define amdgpu_kernel void @test_fmax_legacy_ogt_v1f32(<1 x float> addrspace(1)* %out, <1 x float> addrspace(1)* %in) #0 {
@@ -132,12 +167,24 @@ define amdgpu_kernel void @test_fmax_legacy_ogt_v1f32(<1 x float> addrspace(1)*
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_ogt_v3f32:
-; GCN-SAFE: v_max_legacy_f32_e32
-; GCN-SAFE: v_max_legacy_f32_e32
-; GCN-SAFE: v_max_legacy_f32_e32
+; SI-SAFE: v_max_legacy_f32_e32
+; SI-SAFE: v_max_legacy_f32_e32
+; SI-SAFE: v_max_legacy_f32_e32
+
+; VI-SAFE: v_cmp_gt_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-SAFE: v_cmp_gt_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-SAFE: v_cmp_gt_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-SAFE-NOT: v_cmp
+; VI-SAFE-NOT: v_cndmask
+
 ; GCN-NONAN: v_max_f32_e32
 ; GCN-NONAN: v_max_f32_e32
 ; GCN-NONAN: v_max_f32_e32
+
+; GCN-NOT: v_max
 define amdgpu_kernel void @test_fmax_legacy_ogt_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
   %gep.0 = getelementptr <3 x float>, <3 x float> addrspace(1)* %in, i32 %tid
@@ -153,8 +200,8 @@ define amdgpu_kernel void @test_fmax_legacy_ogt_v3f32(<3 x float> addrspace(1)*
 }
 
 ; FUNC-LABEL: {{^}}test_fmax_legacy_ogt_f32_multi_use:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
 ; GCN-NOT: v_max_
 ; GCN: v_cmp_gt_f32
 ; GCN-NEXT: v_cndmask_b32

diff --git a/llvm/test/CodeGen/AMDGPU/fmaxnum.ll b/llvm/test/CodeGen/AMDGPU/fmaxnum.ll
@@ -1,14 +1,26 @@
 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
 
-; GCN-LABEL: {{^}}test_fmax_f32:
-; GCN: v_max_f32_e32
-define amdgpu_kernel void @test_fmax_f32(float addrspace(1)* %out, float %a, float %b) #0 {
-  %val = call float @llvm.maxnum.f32(float %a, float %b)
+; GCN-LABEL: {{^}}test_fmax_f32_ieee_mode_on:
+; GCN: v_mul_f32_e64 [[QUIET0:v[0-9]+]], 1.0, s{{[0-9]+}}
+; GCN: v_mul_f32_e64 [[QUIET1:v[0-9]+]], 1.0, s{{[0-9]+}}
+; GCN: v_max_f32_e32 [[RESULT:v[0-9]+]], [[QUIET1]], [[QUIET0]]
+; GCN-NOT: [[RESULT]]
+; GCN: buffer_store_dword [[RESULT]]
+define amdgpu_kernel void @test_fmax_f32_ieee_mode_on(float addrspace(1)* %out, float %a, float %b) #0 {
+  %val = call float @llvm.maxnum.f32(float %a, float %b) #1
   store float %val, float addrspace(1)* %out, align 4
   ret void
 }
 
+; GCN-LABEL: {{^}}test_fmax_f32_ieee_mode_off:
+; GCN: v_max_f32_e32 v0, v0, v1
+; GCN-NEXT: ; return
+define amdgpu_ps float @test_fmax_f32_ieee_mode_off(float %a, float %b) #0 {
+  %val = call float @llvm.maxnum.f32(float %a, float %b) #1
+  ret float %val
+}
+
 ; GCN-LABEL: {{^}}test_fmax_v2f32:
 ; GCN: v_max_f32_e32
 ; GCN: v_max_f32_e32
@@ -158,38 +170,34 @@ define amdgpu_kernel void @constant_fold_fmax_f32_n0_n0(float addrspace(1)* %out
   ret void
 }
 
-; GCN-LABEL: {{^}}fmax_var_immediate_f32:
+; GCN-LABEL: {{^}}fmax_var_immediate_f32_no_ieee:
 ; GCN: v_max_f32_e64 {{v[0-9]+}}, {{s[0-9]+}}, 2.0
-define amdgpu_kernel void @fmax_var_immediate_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.maxnum.f32(float %a, float 2.0)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmax_var_immediate_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.maxnum.f32(float %a, float 2.0) #0
+  ret float %val
 }
 
-; GCN-LABEL: {{^}}fmax_immediate_var_f32:
+; GCN-LABEL: {{^}}fmax_immediate_var_f32_no_ieee:
 ; GCN: v_max_f32_e64 {{v[0-9]+}}, {{s[0-9]+}}, 2.0
-define amdgpu_kernel void @fmax_immediate_var_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.maxnum.f32(float 2.0, float %a)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmax_immediate_var_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.maxnum.f32(float 2.0, float %a) #0
+  ret float %val
 }
 
-; GCN-LABEL: {{^}}fmax_var_literal_f32:
+; GCN-LABEL: {{^}}fmax_var_literal_f32_no_ieee:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x42c60000
 ; GCN: v_max_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, [[REG]]
-define amdgpu_kernel void @fmax_var_literal_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.maxnum.f32(float %a, float 99.0)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmax_var_literal_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.maxnum.f32(float %a, float 99.0) #0
+  ret float %val
 }
 
-; GCN-LABEL: {{^}}fmax_literal_var_f32:
+; GCN-LABEL: {{^}}fmax_literal_var_f32_no_ieee:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x42c60000
 ; GCN: v_max_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, [[REG]]
-define amdgpu_kernel void @fmax_literal_var_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.maxnum.f32(float 99.0, float %a)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmax_literal_var_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.maxnum.f32(float 99.0, float %a) #0
+  ret float %val
 }
 
 ; GCN-LABEL: {{^}}test_func_fmax_v3f32:

diff --git a/llvm/test/CodeGen/AMDGPU/fmin3.ll b/llvm/test/CodeGen/AMDGPU/fmin3.ll
@@ -95,22 +95,26 @@ define amdgpu_kernel void @test_fmin3_olt_1_f16(half addrspace(1)* %out, half ad
 ; SI-NEXT: v_min3_f32
 ; SI-NEXT: v_min3_f32
 
-; VI: v_min_f16_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
-; VI: v_min_f16_e32 v0, v0, v1
-; VI: v_min_f16_sdwa v1, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI: v_min_f16_e32 v0, v2, v0
-; VI: v_min_f16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; VI: v_min_f16_e32 v0, v0, v3
-; VI: v_or_b32_e32 v0, v0, v1
-
-; GFX9: v_pk_min_f16
-; GFX9: v_pk_min_f16
-; GFX9: v_pk_min_f16
-define <2 x half> @no_fmin3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) {
+; VI: s_waitcnt
+; VI-NEXT: v_min_f16_sdwa v4, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-NEXT: v_min_f16_e32 v0, v0, v1
+; VI-NEXT: v_min_f16_sdwa v1, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-NEXT: v_min_f16_e32 v0, v2, v0
+; VI-NEXT: v_min_f16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; VI-NEXT: v_min_f16_e32 v0, v0, v3
+; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT: s_setpc_b64
+
+; GFX9: s_waitcnt
+; GFX9-NEXT: v_pk_min_f16 v0, v0, v1
+; GFX9-NEXT: v_pk_min_f16 v0, v2, v0
+; GFX9-NEXT: v_pk_min_f16 v0, v0, v3
+; GFX9-NEXT: s_setpc_b64
+define <2 x half> @no_fmin3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) #2 {
 entry:
-  %min = tail call fast <2 x half> @llvm.minnum.v2f16(<2 x half> %a, <2 x half> %b)
-  %min1 = tail call fast <2 x half> @llvm.minnum.v2f16(<2 x half> %c, <2 x half> %min)
-  %res = tail call fast <2 x half> @llvm.minnum.v2f16(<2 x half> %min1, <2 x half> %d)
+  %min = call <2 x half> @llvm.minnum.v2f16(<2 x half> %a, <2 x half> %b)
+  %min1 = call <2 x half> @llvm.minnum.v2f16(<2 x half> %c, <2 x half> %min)
+  %res = call <2 x half> @llvm.minnum.v2f16(<2 x half> %min1, <2 x half> %d)
   ret <2 x half> %res
 }
 
@@ -121,3 +125,4 @@ declare <2 x half> @llvm.minnum.v2f16(<2 x half>, <2 x half>)
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone speculatable }
+attributes #2 = { nounwind "no-nans-fp-math"="true" }
diff --git a/llvm/test/CodeGen/AMDGPU/fmin_fmax_legacy.amdgcn.ll b/llvm/test/CodeGen/AMDGPU/fmin_fmax_legacy.amdgcn.ll
@@ -1,9 +1,19 @@
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN-SAFE -check-prefix=GCN %s
-; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN-NONAN -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI-SAFE,GCN,SI %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI-NONAN,GCN,SI %s
+
+; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI-SAFE,GCN,VI %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI-NONAN,GCN,VI %s
 
 ; GCN-LABEL: {{^}}min_fneg_select_regression_0:
-; GCN-SAFE: v_max_legacy_f32_e64 [[MIN:v[0-9]+]], -1.0, -v0
-; GCN-NONAN: v_max_f32_e64 v{{[0-9]+}}, -v0, -1.0
+; GCN-NOT: v_mul
+
+; SI-SAFE: v_max_legacy_f32_e64 [[MIN:v[0-9]+]], -1.0, -v0
+
+; VI-SAFE: v_cmp_nle_f32_e32 vcc, 1.0, v0
+; VI-SAFE-NEXT: v_cndmask_b32_e32 v0, 1.0, v0, vcc
+; VI-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
+
+; GCN-NONAN: v_max_f32_e64 v0, -v0, -1.0
 define amdgpu_ps float @min_fneg_select_regression_0(float %a, float %b) #0 {
   %fneg.a = fsub float -0.0, %a
   %cmp.a = fcmp ult float %a, 1.0
@@ -12,7 +22,14 @@ define amdgpu_ps float @min_fneg_select_regression_0(float %a, float %b) #0 {
 }
 
 ; GCN-LABEL: {{^}}min_fneg_select_regression_posk_0:
-; GCN-SAFE: v_max_legacy_f32_e64 [[MIN:v[0-9]+]], 1.0, -v0
+; GCN-NOT: v_mul
+
+; SI-SAFE: v_max_legacy_f32_e64 [[MIN:v[0-9]+]], 1.0, -v0
+
+; VI-SAFE: v_cmp_nle_f32_e32 vcc, -1.0, v0
+; VI-SAFE-NEXT: v_cndmask_b32_e32 v0, -1.0, v0, vcc
+; VI-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
+
 ; GCN-NONAN: v_max_f32_e64 v{{[0-9]+}}, -v0, 1.0
 define amdgpu_ps float @min_fneg_select_regression_posk_0(float %a, float %b) #0 {
   %fneg.a = fsub float -0.0, %a
@@ -22,19 +39,33 @@ define amdgpu_ps float @min_fneg_select_regression_posk_0(float %a, float %b) #0
 }
 
 ; GCN-LABEL: {{^}}max_fneg_select_regression_0:
-; GCN-SAFE: v_min_legacy_f32_e64 [[MIN:v[0-9]+]], -1.0, -v0
+; GCN-NOT: v_mul
+
+; SI-SAFE: v_min_legacy_f32_e64 [[MIN:v[0-9]+]], -1.0, -v0
+
+; VI-SAFE: v_cmp_nge_f32_e32 vcc, 1.0, v0
+; VI-SAFE-NEXT: v_cndmask_b32_e32 v0, 1.0, v0, vcc
+; VI-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
+
 ; GCN-NONAN: v_min_f32_e64 [[MIN:v[0-9]+]], -v0, -1.0
-define amdgpu_ps float @max_fneg_select_regression_0(float %a, float %b) #0 {
+define amdgpu_ps float @max_fneg_select_regression_0(float %a) #0 {
   %fneg.a = fsub float -0.0, %a
   %cmp.a = fcmp ugt float %a, 1.0
   %min.a = select i1 %cmp.a, float %fneg.a, float -1.0
   ret float %min.a
 }
 
 ; GCN-LABEL: {{^}}max_fneg_select_regression_posk_0:
-; GCN-SAFE: v_min_legacy_f32_e64 [[MIN:v[0-9]+]], 1.0, -v0
+; GCN-NOT: v_mul
+
+; SI-SAFE: v_min_legacy_f32_e64 [[MIN:v[0-9]+]], 1.0, -v0
+
+; VI-SAFE: v_cmp_nge_f32_e32 vcc, -1.0, v0
+; VI-SAFE-NEXT: v_cndmask_b32_e32 v0, -1.0, v0, vcc
+; VI-SAFE-NEXT: v_xor_b32_e32 v0, 0x80000000, v0
+
 ; GCN-NONAN: v_min_f32_e64 [[MIN:v[0-9]+]], -v0, 1.0
-define amdgpu_ps float @max_fneg_select_regression_posk_0(float %a, float %b) #0 {
+define amdgpu_ps float @max_fneg_select_regression_posk_0(float %a) #0 {
   %fneg.a = fsub float -0.0, %a
   %cmp.a = fcmp ugt float %a, -1.0
   %min.a = select i1 %cmp.a, float %fneg.a, float 1.0

diff --git a/llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll b/llvm/test/CodeGen/AMDGPU/fmin_legacy.f16.ll
@@ -98,7 +98,7 @@ define <2 x half> @test_fmin_legacy_ule_v2f16(<2 x half> %a, <2 x half> %b) #0 {
 ; VI-NNAN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; VI-NNAN-NEXT:    v_min_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_min_f16_e32 v0, v0, v1
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v2
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmin_legacy_ule_v2f16:
@@ -179,7 +179,7 @@ define <3 x half> @test_fmin_legacy_ule_v3f16(<3 x half> %a, <3 x half> %b) #0 {
 ; VI-NNAN-NEXT:    v_min_f16_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_min_f16_e32 v0, v0, v2
 ; VI-NNAN-NEXT:    v_min_f16_e32 v1, v1, v3
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v4
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmin_legacy_ule_v3f16:
@@ -284,8 +284,8 @@ define <4 x half> @test_fmin_legacy_ule_v4f16(<4 x half> %a, <4 x half> %b) #0 {
 ; VI-NNAN-NEXT:    v_min_f16_e32 v1, v1, v3
 ; VI-NNAN-NEXT:    v_min_f16_sdwa v5, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_min_f16_e32 v0, v0, v2
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v5
-; VI-NNAN-NEXT:    v_or_b32_e32 v1, v1, v4
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v1, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmin_legacy_ule_v4f16:
@@ -438,10 +438,10 @@ define <8 x half> @test_fmin_legacy_ule_v8f16(<8 x half> %a, <8 x half> %b) #0 {
 ; VI-NNAN-NEXT:    v_min_f16_e32 v1, v1, v5
 ; VI-NNAN-NEXT:    v_min_f16_sdwa v11, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
 ; VI-NNAN-NEXT:    v_min_f16_e32 v0, v0, v4
-; VI-NNAN-NEXT:    v_or_b32_e32 v0, v0, v11
-; VI-NNAN-NEXT:    v_or_b32_e32 v1, v1, v10
-; VI-NNAN-NEXT:    v_or_b32_e32 v2, v2, v9
-; VI-NNAN-NEXT:    v_or_b32_e32 v3, v3, v8
+; VI-NNAN-NEXT:    v_or_b32_sdwa v0, v0, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v1, v1, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v2, v2, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NNAN-NEXT:    v_or_b32_sdwa v3, v3, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; VI-NNAN-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SI-SAFE-LABEL: test_fmin_legacy_ule_v8f16:

diff --git a/llvm/test/CodeGen/AMDGPU/fmin_legacy.ll b/llvm/test/CodeGen/AMDGPU/fmin_legacy.ll
@@ -1,5 +1,9 @@
-; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN-SAFE -check-prefix=GCN -check-prefix=FUNC %s
-; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN-NONAN -check-prefix=GCN -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN-SAFE,SI-SAFE,GCN,FUNC %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=SI-NONAN,GCN-NONAN,GCN,FUNC %s
+
+; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI-SAFE,GCN-SAFE,GCN,FUNC %s
+; RUN: llc -enable-no-nans-fp-math -enable-no-signed-zeros-fp-math -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI-NONAN,GCN-NONAN,GCN,FUNC %s
+
 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -enable-var-scope -check-prefix=EG -check-prefix=FUNC %s
 
 declare i32 @llvm.r600.read.tidig.x() #1
@@ -10,8 +14,13 @@ declare i32 @llvm.r600.read.tidig.x() #1
 
 ; FUNC-LABEL: {{^}}s_test_fmin_legacy_subreg_inputs_f32:
 ; EG: MIN *
-; GCN-SAFE: v_min_legacy_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}
-; GCN-NONAN: v_min_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}
+; SI-SAFE: v_min_legacy_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}
+
+; SI-NONAN: v_min_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}
+
+; VI-SAFE: v_cmp_nlt_f32_e32 vcc, s{{[0-9]+}}, v{{[0-9]+}}
+
+; VI-NONAN: v_min_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}
 define amdgpu_kernel void @s_test_fmin_legacy_subreg_inputs_f32(float addrspace(1)* %out, <4 x float> %reg0) #0 {
    %r0 = extractelement <4 x float> %reg0, i32 0
    %r1 = extractelement <4 x float> %reg0, i32 1
@@ -22,27 +31,37 @@ define amdgpu_kernel void @s_test_fmin_legacy_subreg_inputs_f32(float addrspace(
 }
 
 ; FUNC-LABEL: {{^}}s_test_fmin_legacy_ule_f32:
-; GCN-DAG: s_load_dwordx2 s{{\[}}[[A:[0-9]+]]:[[B:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; GCN-DAG: s_load_dwordx2 s{{\[}}[[A:[0-9]+]]:[[B:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
 
 ; GCN-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], s[[B]]
 
-; GCN-SAFE: v_min_legacy_f32_e64 {{v[0-9]+}}, [[VB]], s[[A]]
-; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, s[[A]], [[VB]]
+; SI-SAFE: v_min_legacy_f32_e64 {{v[0-9]+}}, [[VB]], s[[A]]
+
+; VI-SAFE: v_mov_b32_e32 [[VA:v[0-9]+]], s[[A]]
+; VI-SAFE: v_cmp_ngt_f32_e32 vcc, s[[A]], [[VB]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[VB]], [[VA]]
 
+; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, s[[A]], [[VB]]
 define amdgpu_kernel void @s_test_fmin_legacy_ule_f32(float addrspace(1)* %out, float %a, float %b) #0 {
   %cmp = fcmp ule float %a, %b
   %val = select i1 %cmp, float %a, float %b
   store float %val, float addrspace(1)* %out, align 4
   ret void
 }
 
+; Nsz also needed
+; FIXME: Should separate tests
 ; GCN-LABEL: {{^}}s_test_fmin_legacy_ule_f32_nnan_src:
-; GCN: s_load_dwordx2 s{{\[}}[[A:[0-9]+]]:[[B:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; GCN: s_load_dwordx2 s{{\[}}[[A:[0-9]+]]:[[B:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0xb|0x2c}}
 
 ; GCN-DAG: v_add_f32_e64 [[ADD_A:v[0-9]+]], s[[A]], 1.0
 ; GCN-DAG: v_add_f32_e64 [[ADD_B:v[0-9]+]], s[[B]], 2.0
 
-; GCN-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[ADD_B]], [[ADD_A]]
+; SI-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[ADD_B]], [[ADD_A]]
+
+; VI-SAFE: v_cmp_ngt_f32_e32 vcc, [[ADD_A]], [[ADD_B]]
+; VI-SAFE: v_cndmask_b32_e32 {{v[0-9]+}}, [[ADD_B]], [[ADD_A]], vcc
+
 ; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, [[ADD_A]], [[ADD_B]]
 define amdgpu_kernel void @s_test_fmin_legacy_ule_f32_nnan_src(float addrspace(1)* %out, float %a, float %b) #0 {
   %a.nnan = fadd nnan float %a, 1.0
@@ -54,9 +73,14 @@ define amdgpu_kernel void @s_test_fmin_legacy_ule_f32_nnan_src(float addrspace(1
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ule_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+
+; VI-SAFE: v_cmp_ngt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 define amdgpu_kernel void @test_fmin_legacy_ule_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
@@ -73,9 +97,14 @@ define amdgpu_kernel void @test_fmin_legacy_ule_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ole_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+
+; VI-SAFE v_cmp_le_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 define amdgpu_kernel void @test_fmin_legacy_ole_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
@@ -92,9 +121,14 @@ define amdgpu_kernel void @test_fmin_legacy_ole_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_olt_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
+
+; VI-SAFE v_cmp_lt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 define amdgpu_kernel void @test_fmin_legacy_olt_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
@@ -111,9 +145,14 @@ define amdgpu_kernel void @test_fmin_legacy_olt_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ult_f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+
+; VI-SAFE v_cmp_lt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 define amdgpu_kernel void @test_fmin_legacy_ult_f32(float addrspace(1)* %out, float addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
@@ -130,9 +169,14 @@ define amdgpu_kernel void @test_fmin_legacy_ult_f32(float addrspace(1)* %out, fl
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ult_v1f32:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
+
+; SI-SAFE: v_min_legacy_f32_e32 {{v[0-9]+}}, [[B]], [[A]]
+
+; VI-SAFE v_cmp_lt_f32_e32 vcc, [[A]], [[B]]
+; VI-SAFE: v_cndmask_b32_e32 v{{[0-9]+}}, [[B]], [[A]]
+
 ; GCN-NONAN: v_min_f32_e32 {{v[0-9]+}}, [[A]], [[B]]
 define amdgpu_kernel void @test_fmin_legacy_ult_v1f32(<1 x float> addrspace(1)* %out, <1 x float> addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
@@ -149,10 +193,15 @@ define amdgpu_kernel void @test_fmin_legacy_ult_v1f32(<1 x float> addrspace(1)*
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ult_v2f32:
-; GCN: buffer_load_dwordx2
-; GCN: buffer_load_dwordx2
-; GCN-SAFE: v_min_legacy_f32_e32
-; GCN-SAFE: v_min_legacy_f32_e32
+; GCN: {{buffer|flat}}_load_dwordx2
+; GCN: {{buffer|flat}}_load_dwordx2
+; SI-SAFE: v_min_legacy_f32_e32
+; SI-SAFE: v_min_legacy_f32_e32
+
+; VI-SAFE v_cmp_lt_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-SAFE v_cmp_lt_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
 
 ; GCN-NONAN: v_min_f32_e32
 ; GCN-NONAN: v_min_f32_e32
@@ -171,13 +220,24 @@ define amdgpu_kernel void @test_fmin_legacy_ult_v2f32(<2 x float> addrspace(1)*
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ult_v3f32:
-; GCN-SAFE: v_min_legacy_f32_e32
-; GCN-SAFE: v_min_legacy_f32_e32
-; GCN-SAFE: v_min_legacy_f32_e32
+; SI-SAFE: v_min_legacy_f32_e32
+; SI-SAFE: v_min_legacy_f32_e32
+; SI-SAFE: v_min_legacy_f32_e32
+; SI-SAFE-NOT: v_min_
+
+; VI-SAFE: v_cmp_nge_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-SAFE: v_cmp_nge_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-SAFE: v_cmp_nge_f32_e32
+; VI-SAFE: v_cndmask_b32_e32
+; VI-NOT: v_cmp
+; VI-NOT: v_cndmask
 
 ; GCN-NONAN: v_min_f32_e32
 ; GCN-NONAN: v_min_f32_e32
 ; GCN-NONAN: v_min_f32_e32
+; GCN-NONAN-NOT: v_min_
 define amdgpu_kernel void @test_fmin_legacy_ult_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %in) #0 {
   %tid = call i32 @llvm.r600.read.tidig.x() #1
   %gep.0 = getelementptr <3 x float>, <3 x float> addrspace(1)* %in, i32 %tid
@@ -193,8 +253,8 @@ define amdgpu_kernel void @test_fmin_legacy_ult_v3f32(<3 x float> addrspace(1)*
 }
 
 ; FUNC-LABEL: {{^}}test_fmin_legacy_ole_f32_multi_use:
-; GCN: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
-; GCN: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: {{buffer|flat}}_load_dword [[B:v[0-9]+]]
 ; GCN-NOT: v_min
 ; GCN: v_cmp_le_f32
 ; GCN-NEXT: v_cndmask_b32

diff --git a/llvm/test/CodeGen/AMDGPU/fminnum.f64.ll b/llvm/test/CodeGen/AMDGPU/fminnum.f64.ll
@@ -7,15 +7,35 @@ declare <4 x double> @llvm.minnum.v4f64(<4 x double>, <4 x double>) #0
 declare <8 x double> @llvm.minnum.v8f64(<8 x double>, <8 x double>) #0
 declare <16 x double> @llvm.minnum.v16f64(<16 x double>, <16 x double>) #0
 
-; FUNC-LABEL: @test_fmin_f64
-; SI: v_min_f64
-define amdgpu_kernel void @test_fmin_f64(double addrspace(1)* %out, double %a, double %b) nounwind {
+; FUNC-LABEL: {{^}}test_fmin_f64_ieee:
+; SI: s_load_dwordx2 [[A:s\[[0-9]+:[0-9]+\]]]
+; SI: s_load_dwordx2 [[B:s\[[0-9]+:[0-9]+\]]]
+; SI-DAG: v_max_f64 [[QUIETA:v\[[0-9]+:[0-9]+\]]], [[A]], [[A]]
+; SI-DAG: v_max_f64 [[QUIETB:v\[[0-9]+:[0-9]+\]]], [[B]], [[B]]
+; SI: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[QUIETA]], [[QUIETB]]
+define amdgpu_kernel void @test_fmin_f64_ieee([8 x i32], double %a, [8 x i32], double %b) nounwind {
+  %val = call double @llvm.minnum.f64(double %a, double %b) #0
+  store double %val, double addrspace(1)* undef, align 8
+  ret void
+}
+
+; FUNC-LABEL: {{^}}test_fmin_f64_no_ieee:
+; SI: ds_read_b64 [[VAL0:v\[[0-9]+:[0-9]+\]]]
+; SI: ds_read_b64 [[VAL1:v\[[0-9]+:[0-9]+\]]]
+; SI-NOT: [[VAL0]]
+; SI-NOT: [[VAL1]]
+; SI: v_min_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[VAL0]], [[VAL1]]
+; SI-NOT: [[RESULT]]
+; SI: ds_write_b64 v{{[0-9]+}}, [[RESULT]]
+define amdgpu_ps void @test_fmin_f64_no_ieee() nounwind {
+  %a = load volatile double, double addrspace(3)* undef
+  %b = load volatile double, double addrspace(3)* undef
   %val = call double @llvm.minnum.f64(double %a, double %b) #0
-  store double %val, double addrspace(1)* %out, align 8
+  store volatile double %val, double addrspace(3)* undef
   ret void
 }
 
-; FUNC-LABEL: @test_fmin_v2f64
+; FUNC-LABEL: {{^}}test_fmin_v2f64:
 ; SI: v_min_f64
 ; SI: v_min_f64
 define amdgpu_kernel void @test_fmin_v2f64(<2 x double> addrspace(1)* %out, <2 x double> %a, <2 x double> %b) nounwind {
@@ -24,7 +44,7 @@ define amdgpu_kernel void @test_fmin_v2f64(<2 x double> addrspace(1)* %out, <2 x
   ret void
 }
 
-; FUNC-LABEL: @test_fmin_v4f64
+; FUNC-LABEL: {{^}}test_fmin_v4f64:
 ; SI: v_min_f64
 ; SI: v_min_f64
 ; SI: v_min_f64
@@ -35,7 +55,7 @@ define amdgpu_kernel void @test_fmin_v4f64(<4 x double> addrspace(1)* %out, <4 x
   ret void
 }
 
-; FUNC-LABEL: @test_fmin_v8f64
+; FUNC-LABEL: {{^}}test_fmin_v8f64:
 ; SI: v_min_f64
 ; SI: v_min_f64
 ; SI: v_min_f64
@@ -50,7 +70,7 @@ define amdgpu_kernel void @test_fmin_v8f64(<8 x double> addrspace(1)* %out, <8 x
   ret void
 }
 
-; FUNC-LABEL: @test_fmin_v16f64
+; FUNC-LABEL: {{^}}test_fmin_v16f64:
 ; SI: v_min_f64
 ; SI: v_min_f64
 ; SI: v_min_f64

diff --git a/llvm/test/CodeGen/AMDGPU/fminnum.ll b/llvm/test/CodeGen/AMDGPU/fminnum.ll
@@ -1,14 +1,45 @@
 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
 
-; GCN-LABEL: {{^}}test_fmin_f32:
-; GCN: v_min_f32_e32
-define amdgpu_kernel void @test_fmin_f32(float addrspace(1)* %out, float %a, float %b) #0 {
-  %val = call float @llvm.minnum.f32(float %a, float %b)
+; GCN-LABEL: {{^}}test_fmin_f32_ieee_mode_on:
+; GCN: v_mul_f32_e64 [[QUIET0:v[0-9]+]], 1.0, s{{[0-9]+}}
+; GCN: v_mul_f32_e64 [[QUIET1:v[0-9]+]], 1.0, s{{[0-9]+}}
+; GCN: v_min_f32_e32 [[RESULT:v[0-9]+]], [[QUIET1]], [[QUIET0]]
+; GCN-NOT: [[RESULT]]
+; GCN: buffer_store_dword [[RESULT]]
+define amdgpu_kernel void @test_fmin_f32_ieee_mode_on(float addrspace(1)* %out, float %a, float %b) #0 {
+  %val = call float @llvm.minnum.f32(float %a, float %b) #1
   store float %val, float addrspace(1)* %out, align 4
   ret void
 }
 
+; GCN-LABEL: {{^}}test_fmin_nnan_f32_ieee_mode_on:
+; GCN: s_waitcnt
+; GCN-NEXT: v_min_f32_e32 v0, v0, v1
+; GCN-NEXT: s_setpc_b64
+define float @test_fmin_nnan_f32_ieee_mode_on(float %a, float %b) #0 {
+  %val = call nnan float @llvm.minnum.f32(float %a, float %b) #1
+  ret float %val
+}
+
+; GCN-LABEL: {{^}}test_fmin_nnan_f32_ieee_mode_off:
+; GCN-NOT: v0
+; GCN-NOT: v1
+; GCN: v_min_f32_e32 v0, v0, v1
+; GCN-NEXT: ; return
+define amdgpu_ps float @test_fmin_nnan_f32_ieee_mode_off(float %a, float %b) #0 {
+  %val = call nnan float @llvm.minnum.f32(float %a, float %b) #1
+  ret float %val
+}
+
+; GCN-LABEL: {{^}}test_fmin_f32_ieee_mode_off:
+; GCN: v_min_f32_e32 v0, v0, v1
+; GCN-NEXT: ; return
+define amdgpu_ps float @test_fmin_f32_ieee_mode_off(float %a, float %b) #0 {
+  %val = call float @llvm.minnum.f32(float %a, float %b) #1
+  ret float %val
+}
+
 ; GCN-LABEL: {{^}}test_fmin_v2f32:
 ; GCN: v_min_f32_e32
 ; GCN: v_min_f32_e32
@@ -147,38 +178,34 @@ define amdgpu_kernel void @constant_fold_fmin_f32_n0_n0(float addrspace(1)* %out
   ret void
 }
 
-; GCN-LABEL: {{^}}fmin_var_immediate_f32:
-; GCN: v_min_f32_e64 {{v[0-9]+}}, {{s[0-9]+}}, 2.0
-define amdgpu_kernel void @fmin_var_immediate_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.minnum.f32(float %a, float 2.0)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+; GCN-LABEL: {{^}}fmin_var_immediate_f32_no_ieee:
+; GCN: v_min_f32_e32 v0, 2.0, v0
+define amdgpu_ps float @fmin_var_immediate_f32_no_ieee(float %a) #0 {
+  %val = call float @llvm.minnum.f32(float %a, float 2.0) #1
+  ret float %val
 }
 
-; GCN-LABEL: {{^}}fmin_immediate_var_f32:
+; GCN-LABEL: {{^}}fmin_immediate_var_f32_no_ieee:
 ; GCN: v_min_f32_e64 {{v[0-9]+}}, {{s[0-9]+}}, 2.0
-define amdgpu_kernel void @fmin_immediate_var_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.minnum.f32(float 2.0, float %a)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmin_immediate_var_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.minnum.f32(float 2.0, float %a) #1
+  ret float %val
 }
 
-; GCN-LABEL: {{^}}fmin_var_literal_f32:
+; GCN-LABEL: {{^}}fmin_var_literal_f32_no_ieee:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x42c60000
 ; GCN: v_min_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, [[REG]]
-define amdgpu_kernel void @fmin_var_literal_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.minnum.f32(float %a, float 99.0)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmin_var_literal_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.minnum.f32(float %a, float 99.0) #1
+  ret float %val
 }
 
-; GCN-LABEL: {{^}}fmin_literal_var_f32:
+; GCN-LABEL: {{^}}fmin_literal_var_f32_no_ieee:
 ; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x42c60000
 ; GCN: v_min_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}, [[REG]]
-define amdgpu_kernel void @fmin_literal_var_f32(float addrspace(1)* %out, float %a) #0 {
-  %val = call float @llvm.minnum.f32(float 99.0, float %a)
-  store float %val, float addrspace(1)* %out, align 4
-  ret void
+define amdgpu_ps float @fmin_literal_var_f32_no_ieee(float inreg %a) #0 {
+  %val = call float @llvm.minnum.f32(float 99.0, float %a) #1
+  ret float %val
 }
 
 ; GCN-LABEL: {{^}}test_func_fmin_v3f32:

diff --git a/llvm/test/CodeGen/AMDGPU/fneg-combines.ll b/llvm/test/CodeGen/AMDGPU/fneg-combines.ll
diff --git a/llvm/test/CodeGen/AMDGPU/known-never-snan.ll b/llvm/test/CodeGen/AMDGPU/known-never-snan.ll
@@ -99,8 +99,7 @@ define float @v_test_known_not_snan_minnum_input_fmed3_r_i_i_f32(float %a, float
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
 ; GCN-NEXT:    v_min_f32_e32 v0, v0, v1
-; GCN-NEXT:    v_max_f32_e32 v0, 2.0, v0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %a.nnan.add = fdiv nnan float 1.0, %a
   %b.nnan.add = fadd nnan float %b, 1.0
@@ -110,14 +109,46 @@ define float @v_test_known_not_snan_minnum_input_fmed3_r_i_i_f32(float %a, float
   ret float %med
 }
 
+define float @v_test_known_not_minnum_maybe_nan_src0_input_fmed3_r_i_i_f32(float %a, float %b) #0 {
+; GCN-LABEL: v_test_known_not_minnum_maybe_nan_src0_input_fmed3_r_i_i_f32:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
+; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
+; GCN-NEXT:    v_min_f32_e32 v0, v0, v1
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %b.nsnan = fadd float %b, 1.0
+  %known.not.snan = call float @llvm.minnum.f32(float %a, float %b.nsnan)
+  %max = call float @llvm.maxnum.f32(float %known.not.snan, float 2.0)
+  %med = call float @llvm.minnum.f32(float %max, float 4.0)
+  ret float %med
+}
+
+define float @v_test_known_not_minnum_maybe_nan_src1_input_fmed3_r_i_i_f32(float %a, float %b) #0 {
+; GCN-LABEL: v_test_known_not_minnum_maybe_nan_src1_input_fmed3_r_i_i_f32:
+; GCN:       ; %bb.0:
+; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GCN-NEXT:    v_add_f32_e32 v0, 1.0, v0
+; GCN-NEXT:    v_mul_f32_e32 v1, 1.0, v1
+; GCN-NEXT:    v_min_f32_e32 v0, v0, v1
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GCN-NEXT:    s_setpc_b64 s[30:31]
+  %a.nsnan = fadd float %a, 1.0
+  %known.not.snan = call float @llvm.minnum.f32(float %a.nsnan, float %b)
+  %max = call float @llvm.maxnum.f32(float %known.not.snan, float 2.0)
+  %med = call float @llvm.minnum.f32(float %max, float 4.0)
+  ret float %med
+}
+
 define float @v_minnum_possible_nan_lhs_input_fmed3_r_i_i_f32(float %a, float %b) #0 {
 ; GCN-LABEL: v_minnum_possible_nan_lhs_input_fmed3_r_i_i_f32:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
+; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
 ; GCN-NEXT:    v_min_f32_e32 v0, v0, v1
-; GCN-NEXT:    v_max_f32_e32 v0, 2.0, v0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %b.nnan.add = fadd nnan float %b, 1.0
   %known.not.snan = call float @llvm.minnum.f32(float %a, float %b.nnan.add)
@@ -131,9 +162,9 @@ define float @v_minnum_possible_nan_rhs_input_fmed3_r_i_i_f32(float %a, float %b
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
+; GCN-NEXT:    v_mul_f32_e32 v1, 1.0, v1
 ; GCN-NEXT:    v_min_f32_e32 v0, v0, v1
-; GCN-NEXT:    v_max_f32_e32 v0, 2.0, v0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %a.nnan.add = fdiv nnan float 1.0, %a
   %known.not.snan = call float @llvm.minnum.f32(float %a.nnan.add, float %b)
@@ -148,8 +179,8 @@ define float @v_test_known_not_snan_maxnum_input_fmed3_r_i_i_f32(float %a, float
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT:    v_max3_f32 v0, v0, v1, 2.0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_max_f32_e32 v0, v0, v1
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %a.nnan.add = fdiv nnan float 1.0, %a
   %b.nnan.add = fadd nnan float %b, 1.0
@@ -164,8 +195,9 @@ define float @v_maxnum_possible_nan_lhs_input_fmed3_r_i_i_f32(float %a, float %b
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
-; GCN-NEXT:    v_max3_f32 v0, v0, v1, 2.0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
+; GCN-NEXT:    v_max_f32_e32 v0, v0, v1
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %b.nnan.add = fadd nnan float %b, 1.0
   %known.not.snan = call float @llvm.maxnum.f32(float %a, float %b.nnan.add)
@@ -179,8 +211,9 @@ define float @v_maxnum_possible_nan_rhs_input_fmed3_r_i_i_f32(float %a, float %b
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
-; GCN-NEXT:    v_max3_f32 v0, v0, v1, 2.0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_mul_f32_e32 v1, 1.0, v1
+; GCN-NEXT:    v_max_f32_e32 v0, v0, v1
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %a.nnan.add = fdiv nnan float 1.0, %a
   %known.not.snan = call float @llvm.maxnum.f32(float %a.nnan.add, float %b)
@@ -215,8 +248,8 @@ define float @v_select_possible_nan_lhs_input_fmed3_r_i_i_f32(float %a, float %b
 ; GCN-NEXT:    v_add_f32_e32 v1, 1.0, v1
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
-; GCN-NEXT:    v_max_f32_e32 v0, 2.0, v0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %b.nnan.add = fadd nnan float %b, 1.0
   %cmp = icmp eq i32 %c, 0
@@ -233,8 +266,8 @@ define float @v_select_possible_nan_rhs_input_fmed3_r_i_i_f32(float %a, float %b
 ; GCN-NEXT:    v_rcp_f32_e32 v0, v0
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v2
 ; GCN-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
-; GCN-NEXT:    v_max_f32_e32 v0, 2.0, v0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %a.nnan.add = fdiv nnan float 1.0, %a
   %cmp = icmp eq i32 %c, 0
@@ -494,6 +527,7 @@ define float @v_test_known_not_snan_fmed3_input_fmed3_r_i_i_f32(float %a, float
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_med3_f32 v0, v0, v1, v2
+; GCN-NEXT:    v_mul_f32_e32 v0, 1.0, v0
 ; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %known.not.snan = call float @llvm.amdgcn.fmed3.f32(float %a, float %b, float %c)
@@ -507,8 +541,7 @@ define float @v_test_known_not_snan_fmin3_input_fmed3_r_i_i_f32(float %a, float
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    v_min3_f32 v0, v0, v1, v2
-; GCN-NEXT:    v_max_f32_e32 v0, 2.0, v0
-; GCN-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GCN-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %min0 = call float @llvm.minnum.f32(float %a, float %b)
   %known.not.snan = call float @llvm.minnum.f32(float %min0, float %c)

diff --git a/llvm/test/CodeGen/AMDGPU/llvm.maxnum.f16.ll b/llvm/test/CodeGen/AMDGPU/llvm.maxnum.f16.ll
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.minnum.f16.ll b/llvm/test/CodeGen/AMDGPU/llvm.minnum.f16.ll
diff --git a/llvm/test/CodeGen/AMDGPU/reduction.ll b/llvm/test/CodeGen/AMDGPU/reduction.ll
@@ -434,12 +434,23 @@ entry:
 }
 
 ; GCN-LABEL: {{^}}reduction_maxnum_v4f16:
-; GFX9:      v_pk_max_f16 [[MAX:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX9-NEXT: v_max_f16_sdwa v{{[0-9]+}}, [[MAX]], [[MAX]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-
-; VI:      v_max_f16_sdwa
-; VI-NEXT: v_max_f16_e32
-; VI-NEXT: v_max_f16_e32
+; GFX9: s_waitcnt
+; GFX9-NEXT: v_pk_max_f16 [[CANON1:v[0-9]+]], v1, v1
+; GFX9-NEXT: v_pk_max_f16 [[CANON0:v[0-9]+]], v0, v0
+; GFX9-NEXT: v_pk_max_f16 [[MAX:v[0-9]+]], [[CANON0]], [[CANON1]]{{$}}
+
+; FIXME: Extra canonicalize leftover
+; GFX9-NEXT: v_max_f16_sdwa [[TMP:v[0-9]+]], [[MAX]], [[MAX]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX9-NEXT: v_max_f16_e32 v0, [[MAX]], [[TMP]]
+
+; VI-DAG: v_max_f16_sdwa [[CANON1:v[0-9]+]], v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_sdwa [[CANON3:v[0-9]+]], v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_e32 [[CANON0:v[0-9]+]], v0, v0
+; VI-DAG: v_max_f16_e32 [[CANON2:v[0-9]+]], v1, v1
+
+; VI-DAG: v_max_f16_e32 [[MAX0:v[0-9]+]], [[CANON1]], [[CANON3]]
+; VI-DAG: v_max_f16_e32 [[MAX1:v[0-9]+]], [[CANON0]], [[CANON2]]
+; VI: v_max_f16_e32 v0, [[MAX1]], [[MAX0]]
 define half @reduction_maxnum_v4f16(<4 x half> %vec4) {
 entry:
   %rdx.shuf = shufflevector <4 x half> %vec4, <4 x half> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
@@ -451,12 +462,24 @@ entry:
 }
 
 ; GCN-LABEL: {{^}}reduction_minnum_v4f16:
-; GFX9:      v_pk_min_f16 [[MIN:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX9-NEXT: v_min_f16_sdwa v{{[0-9]+}}, [[MIN]], [[MIN]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9: s_waitcnt
+; GFX9-NEXT: v_pk_max_f16 [[CANON1:v[0-9]+]], v1, v1
+; GFX9-NEXT: v_pk_max_f16 [[CANON0:v[0-9]+]], v0, v0
+; GFX9-NEXT: v_pk_min_f16 [[MIN:v[0-9]+]], [[CANON0]], [[CANON1]]{{$}}
+
+; FIXME: Extra canonicalize leftover
+; GFX9-NEXT: v_max_f16_sdwa [[TMP:v[0-9]+]], [[MIN]], [[MIN]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX9-NEXT: v_min_f16_e32 v0, [[MIN]], [[TMP]]
 
-; VI:      v_min_f16_sdwa
-; VI-NEXT: v_min_f16_e32
-; VI-NEXT: v_min_f16_e32
+
+; VI-DAG: v_max_f16_sdwa [[CANON1:v[0-9]+]], v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_sdwa [[CANON3:v[0-9]+]], v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_e32 [[CANON0:v[0-9]+]], v0, v0
+; VI-DAG: v_max_f16_e32 [[CANON2:v[0-9]+]], v1, v1
+
+; VI-DAG: v_min_f16_e32 [[MAX0:v[0-9]+]], [[CANON1]], [[CANON3]]
+; VI-DAG: v_min_f16_e32 [[MAX1:v[0-9]+]], [[CANON0]], [[CANON2]]
+; VI: v_min_f16_e32 v0, [[MAX1]], [[MAX0]]
 define half @reduction_minnum_v4f16(<4 x half> %vec4) {
 entry:
   %rdx.shuf = shufflevector <4 x half> %vec4, <4 x half> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
@@ -467,13 +490,36 @@ entry:
   ret half %res
 }
 
+; FIXME: Need to preserve fast math flags when fmaxnum matched
+; directly from the IR to avoid unnecessary quieting.
+
 ; GCN-LABEL: {{^}}reduction_fast_max_pattern_v4f16:
-; GFX9:      v_pk_max_f16 [[MAX:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX9-NEXT: v_max_f16_sdwa v{{[0-9]+}}, [[MAX]], [[MAX]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; XGFX9:      v_pk_max_f16 [[MAX:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; XGFX9-NEXT: v_max_f16_sdwa v{{[0-9]+}}, [[MAX]], [[MAX]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; XVI: s_waitcnt
+; XVI-NEXT: v_max_f16_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; XVI-NEXT: v_max_f16_e32 v0, v0, v1
+; XVI-NEXT: v_max_f16_e32 v0, v0, v2
+; XVI-NEXT: s_setpc_b64
 
-; VI:      v_max_f16_sdwa
-; VI-NEXT: v_max_f16_e32
-; VI-NEXT: v_max_f16_e32
+; GFX9: s_waitcnt
+; GFX9-NEXT: v_pk_max_f16 [[CANON1:v[0-9]+]], v1, v1
+; GFX9-NEXT: v_pk_max_f16 [[CANON0:v[0-9]+]], v0, v0
+; GFX9-NEXT: v_pk_max_f16 [[MAX:v[0-9]+]], [[CANON0]], [[CANON1]]{{$}}
+
+; FIXME: Extra canonicalize leftover
+; GFX9-NEXT: v_max_f16_sdwa [[TMP:v[0-9]+]], [[MAX]], [[MAX]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX9-NEXT: v_max_f16_e32 v0, [[MAX]], [[TMP]]
+
+; VI-DAG: v_max_f16_sdwa [[CANON1:v[0-9]+]], v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_sdwa [[CANON3:v[0-9]+]], v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_e32 [[CANON0:v[0-9]+]], v0, v0
+; VI-DAG: v_max_f16_e32 [[CANON2:v[0-9]+]], v1, v1
+
+; VI-DAG: v_max_f16_e32 [[MAX0:v[0-9]+]], [[CANON1]], [[CANON3]]
+; VI-DAG: v_max_f16_e32 [[MAX1:v[0-9]+]], [[CANON0]], [[CANON2]]
+; VI: v_max_f16_e32 v0, [[MAX1]], [[MAX0]]
 define half @reduction_fast_max_pattern_v4f16(<4 x half> %vec4) {
 entry:
   %rdx.shuf = shufflevector <4 x half> %vec4, <4 x half> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
@@ -486,13 +532,37 @@ entry:
   ret half %res
 }
 
+; FIXME: Need to preserve fast math flags when fmaxnum matched
+; directly from the IR to avoid unnecessary quieting.
+
 ; GCN-LABEL: {{^}}reduction_fast_min_pattern_v4f16:
-; GFX9:      v_pk_min_f16 [[MIN:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
-; GFX9-NEXT: v_min_f16_sdwa v{{[0-9]+}}, [[MIN]], [[MIN]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; XGFX9:      v_pk_min_f16 [[MIN:v[0-9]+]], v{{[0-9]+}}, v{{[0-9]+}}{{$}}
+; XGFX9-NEXT: v_min_f16_sdwa v{{[0-9]+}}, [[MIN]], [[MIN]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+
+; XVI: s_waitcnt
+; XVI-NEXT: v_min_f16_sdwa v2, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; XVI-NEXT: v_min_f16_e32 v0, v0, v1
+; XVI-NEXT: v_min_f16_e32 v0, v0, v2
+; XVI-NEXT: s_setpc_b64
+
+; GFX9: s_waitcnt
+; GFX9-NEXT: v_pk_max_f16 [[CANON1:v[0-9]+]], v1, v1
+; GFX9-NEXT: v_pk_max_f16 [[CANON0:v[0-9]+]], v0, v0
+; GFX9-NEXT: v_pk_min_f16 [[MIN:v[0-9]+]], [[CANON0]], [[CANON1]]{{$}}
+
+; FIXME: Extra canonicalize leftover
+; GFX9-NEXT: v_max_f16_sdwa [[TMP:v[0-9]+]], [[MIN]], [[MIN]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; GFX9-NEXT: v_min_f16_e32 v0, [[MIN]], [[TMP]]
+
+
+; VI-DAG: v_max_f16_sdwa [[CANON1:v[0-9]+]], v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_sdwa [[CANON3:v[0-9]+]], v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
+; VI-DAG: v_max_f16_e32 [[CANON0:v[0-9]+]], v0, v0
+; VI-DAG: v_max_f16_e32 [[CANON2:v[0-9]+]], v1, v1
 
-; VI:      v_min_f16_sdwa
-; VI-NEXT: v_min_f16_e32
-; VI-NEXT: v_min_f16_e32
+; VI-DAG: v_min_f16_e32 [[MAX0:v[0-9]+]], [[CANON1]], [[CANON3]]
+; VI-DAG: v_min_f16_e32 [[MAX1:v[0-9]+]], [[CANON0]], [[CANON2]]
+; VI: v_min_f16_e32 v0, [[MAX1]], [[MAX0]]
 define half @reduction_fast_min_pattern_v4f16(<4 x half> %vec4) {
 entry:
   %rdx.shuf = shufflevector <4 x half> %vec4, <4 x half> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>